alibaba / wan 3.0 r2v

Overview

Wan 3.0 R2V is a reference-to-video model by Alibaba. It generates a new video from your prompt while following the people, style, or motion you supply as references. You can mix reference images, short video clips, and audio to guide identity, pacing, and voice. The output is a single MP4 video at 30 fps, which is useful for character-consistent shots and repeatable creative iterations.

What you can build

• Character-consistent clips that keep the same face, outfit, and style across a scene
• Product demos that keep brand look consistent while changing camera moves and environments
• Motion-guided remakes where a new subject follows the timing and energy of a reference clip
• Dialogue or voice-preserving scenes using reference audio as guidance
• Storyboard-to-shot experiments using several reference images to lock the look

Inputs

• A required written description of the video you want. You can reference your inputs by index, like “Image 1” or “Video 1”. Long prompts are supported up to 20,000 characters.
• An optional set of up to 10 reference images to guide subject identity or visual style. Use JPG, JPEG, PNG (no transparency), BMP, or WEBP. Each image must be 240–8000 px per side, ratio up to 8:1, and up to 20 MB.
• An optional set of up to 5 reference video clips to guide motion, pacing, framing, or edits. Use MP4 or MOV. Each clip must be 1–15 seconds, at least 16 fps, 240–4096 px per side, ratio up to 8:1, and up to 100 MB. Total reference video time across all clips must stay within 15 seconds.
• An optional set of up to 5 reference audio clips to guide voice, timing, or sound style. Use WAV or MP3. Each clip must be 1–15 seconds and up to 15 MB. Total reference audio time across all clips must stay within 15 seconds.
• A target duration for the generated video, from 2 to 30 seconds.
• A resolution tier for the output video: 480p, 720p, or 1080p.
• An optional aspect ratio choice. “Adaptive” lets the model pick a ratio that matches your intent and references. Fixed ratios include 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16.
• An optional switch to include or omit an audio track in the output.
• An optional switch that rewrites and expands your prompt for more detail. This often helps when your prompt is short.
• An optional random seed (0 to 2,147,483,647) to make runs more repeatable.

Outputs

• One MP4 video.
• The video is generated at 30 fps.
• The platform typically returns basic generation metadata alongside the video, such as the seed used and the measured duration.

Limitations

• If you include reference video, the combined time of reference video plus generated output must stay within 30 seconds.
• Reference indexing matters. “Image 1” maps to the first reference image you provide. Videos and images use separate numbering.
• Identity and style can drift when reference images are low-resolution, heavily compressed, overfiltered, or poorly lit.
• Motion guidance weakens if reference videos have low frame rate, fast cuts, or heavy motion blur.
• Audio guidance works best with clean speech or clean isolated sound. Noisy mixes reduce control.

Safety & compliance

• Don’t generate illegal content or content that promotes harm.
• Don’t create sexual content involving minors or any exploitative content.
• Don’t generate hateful or harassing content targeting protected groups.
• Don’t impersonate real people or clone a voice without permission.
• You’re responsible for rights to any reference images, videos, or audio you upload, including consent and IP rights.